The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires downloading a project from an external GitHub repository (github.com/leigest519/OpenGame) and installing its dependencies using npm install.
[COMMAND_EXECUTION]: Various shell commands are used to interact with the framework, including npm run generate for game creation, npm run bench for evaluation, and npm run serve to launch local web servers.
[REMOTE_CODE_EXECUTION]: The framework generates code (HTML/JS/CSS) based on user prompts and automatically executes it within a headless browser (Playwright) during build-health and usability evaluations. This automated execution cycle can be exploited if malicious prompts result in the generation of harmful scripts.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because user-provided prompts drive the entire logic of code generation, template matching, and iterative debugging repairs.
Ingestion points: Primary natural language prompt provided to OpenGameAgent in SKILL.md.
Boundary markers: None identified in the provided instructions or API examples.
Capability inventory: File system access for writing generated games, network access for LLM/VLM API communication, and automated code execution in a browser environment.
Sanitization: No mechanisms for sanitizing or validating generated code before execution in the benchmarking phase are mentioned.

opengame-agentic-game-creation